One person runs nine platforms: 37Studio puts a content team inside Docker; hold off on auto-replies for now

37Studio is an MIT project that was open-sourced just a few days ago. It runs on your own computer and packs video editing, multi-platform publishing, and a social media inbox into a single Docker container. It works around Claude Code, Anthropic's coding agent that can edit files and run commands from the command line.

What it does is very concrete. Paste in a company website, and it spends five to ten minutes scraping the logo, people, photos, fonts, and color values, then generates a brand library, a design system, and a knowledge.md file. When you create a new project and answer a few questions, Claude Code acts as the editing agent and writes a HyperFrames timeline. HyperFrames is HeyGen's open-source HTML video framework; a video is just a page of HTML with timing attributes. The agent renders still frames to check its own visuals, makes changes if it isn't satisfied, and finally renders frame by frame with headless Chrome and encodes with ffmpeg.

The finished video can be published to nine platforms: Facebook Pages, Instagram, YouTube, TikTok, LinkedIn, X, Pinterest, Bluesky, and Mastodon. Copy is written separately for each platform, and you can publish now or schedule it. Comments and DMs go into one inbox, where Claude drafts replies based on the knowledge base and supports the customer's language.

There's another layer: the Assistant in the sidebar runs the entire studio through MCP. It has 65 tools covering companies, campaigns, videos, publishing, Autopilot, inbox, and costs. Claude Code, Claude Desktop, and Cursor can all connect. MCP is the Model Context Protocol; put simply, it gives an AI agent a list of tools and lets it call them according to that list.

Costs are transparent. Every project, campaign, and company shows how much AI has spent, in USD by default, but it can be changed to your currency. The software itself is MIT and has no monthly fee. You pay the AI bills yourself to Anthropic, Google, and ElevenLabs.

But a few things need to be made clear.

It's only a few days old and has 5 stars on GitHub. One auto-reply option has Claude reply by itself without a human clicking send. In Docker, the launcher gives the agent bypass permissions, meaning any command can run inside the container. Anyone who can log in can make the agent read all companies' files under /data and connect to the internet. Cloud deployment files also default to bypass, protected only by an HTTP basic auth password.

Unreviewed platform apps also have limits. YouTube uploads are private by default; TikTok can only post 'Only me' or save drafts; new Pinterest apps can only use sandbox mode. Threads, Facebook Groups and personal profiles, Reddit, and Snapchat are not on the supported list because the platform APIs do not allow it.

In the replies, someone gave a four-step trial order: first turn off auto-replies and leave only drafts, then click send yourself; use a secondary account, not your main account; set a spending cap on the AI key; and clearly understand what the Docker launcher can touch. If the first step fails, stop. This order is more practical than any feature list.

Another reply put it more precisely: after a video is finished, you can check it slowly; once a comment or DM is sent, it is already part of the conversation. Video has room for error between rendering and publishing; social media replies have room for error before you click send. 37Studio builds 'check its own frames' into the agent workflow, which is already more QA than many brand accounts do. But comments and DMs are a different kind of work—they face real people. When the knowledge base can't answer, the agent says, 'A team member will get back to you.' Once that sentence is sent, someone on the team has to actually follow up.

So the dividing line for this tool isn't whether it can make videos; it's which part you hand over to automation. Autopilot lets each video be generated three hours in advance and wait for a one-click approval, and that design is restrained. Fully automatic inbox replies are another matter. The click it saves is exactly the click where responsibility shifts.

If you want to keep your content workflow in your own hands, 37Studio is worth a look. It puts the brand library, rendering, publishing, inbox, and costs in one local container, with the keys in your hands and the bills in your hands too. If you're looking for a tool that is fully managed as soon as you open it, it can't offer that now, and a project that's only a few days old can't either.

Before using it, do four things first: turn off auto-replies, use a secondary account, set an AI spending cap, and confirm the scope of Docker permissions. After those four things are done, then talk about whether it can save you a content team.